the computer journal 2009 greenspan 43 63

Upload: a-d-prasad

Post on 03-Apr-2018

224 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    1/21

    Super-Resolution in Medical Imaging

    HAYIT GREENSPAN*

    Faculty of Engineering, Biomedical Engineering Department, Tel-Aviv University, Tel-Aviv, Israel

    *Corresponding author: [email protected]

    This paper provides an overview on super-resolution (SR) research in medical imaging

    applications. Many imaging modalities exist. Some provide anatomical information and reveal

    information about the structure of the human body, and others provide functional information,

    locations of activity for specific activities and specified tasks. Each imaging system has a character-

    istic resolution, which is determined based on physical constraints of the system detectors that are in

    turn tuned to signal-to-noise and timing considerations. A common goal across systems is to

    increase the resolution, and as much as possible achieve true isotropic 3-D imaging. SR technology

    can serve to advance this goal. Research on SR in key medical imaging modalities, including MRI,

    fMRI and PET, has started to emerge in recent years and is reviewed herein. The algorithms used

    are mostly based on standard SR algorithms. Results demonstrate the potential in introducing SR

    techniques into practical medical applications.

    Keywords: medical imaging; super-resolution; MRI; PET

    Received 6 August 2006; revised 6 June 2007

    1. INTRODUCTION

    The main goal of medical imaging is to extract a 3-D modeling

    of the human body or specific organs within it. To accomplish

    this goal, various imaging modalities have been developed

    over the years, each based on a particular energy source that

    passes through the body. Many imaging modalities exist,

    with some providing anatomical information and revealinginformation about the structure, and others providing func-

    tional information, e.g. revealing locations of activity within

    the brain for specific activities and specified tasks. The

    medical imaging field is rapidly evolving in increased resol-

    ution machines and advanced content-processing tools [1].

    Medical imaging system developers strive to increase resolu-

    tion since higher resolution is the key to more accurate under-

    standing of the anatomy, it can support early detection of

    abnormalities and can increase the accuracy in the assessment

    of size and morphology of organs and pathologies.

    In recent years, several research groups have started to

    address the goal of resolution augmentation in medical

    imagery as a software post-processing challenge, rather than

    a medical hardware-engineering task. The motivation for

    this initiative emerged following major advances in the

    domains of image and video processing that indicated the

    possibility of augmenting resolution using what are known

    as super-resolution (SR) algorithms. SR deals with the task

    of using several low-resolution (LR) images from a particular

    imaging system to estimate, or reconstruct, the high-resolution

    (HR) source. Each LR input image focuses on a slightly

    shifted field-of-view [or point-of-view (POV)] of the HR

    scene. A variety of reconstruction algorithms have been pro-

    posed in the literature, where the common goal is to estimate

    the HR source as accurately as possible, while minimizing

    noise and preserving important image constraints, including

    image smoothness and more recently, additional prior-

    knowledge about the source.In 2001 and 2002, initial attempts were made to adapt SR

    algorithms from the computer-vision community to medical

    imagery applications. Initial research dealt with the magnetic

    resonance imaging (MRI) modality [2, 3]. Results were

    encouraging and were reproduced around the world within

    the same modalities as well as additional ones, including func-

    tional MRI (fMRI) [4] and positron emission tomography

    (PET) [5, 6]. The goal of the current paper is to review

    several of the key studies that focus on SR algorithms in

    medical applications.

    Following are the key observations that may be made from

    the current overview.

    Research in the field has so far been in what can be defined

    as a hypothesis testing phase: the investigation of a selected

    imaging modality and its evaluation as a candidate for SR,

    and the application of a standard SR algorithm (selected

    from the literature) to a specific medical task.

    Results are encouraging, for the MRI variants, and PET

    modalities.

    Experiments are conducted on phantoms as well as real

    patient data. Results are more substantial on the phantom

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

    # The Author 2008. Published by Oxford University Press on behalf of The British Computer Society. All rights reserved.For Permissions, please email: [email protected]

    Advance Access publication on February 19, 2008 doi:10.1093/comjnl/bxm075

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    2/21

    data. Real data scenarios introduce various artifacts that need

    to be addressed, patient motion difficulties as well as acqui-

    sition time constraints.

    A comparison across several currently published SR var-

    iants algorithmic variants does not reveal major differences.

    Future major challenges in this field include: on the theor-

    etical front, the development of novel SR algorithms thatcombine medical image prior knowledge as regularization

    terms in the SR process. On the applicative front, the shift to

    the clinical settings, an important next step to define the con-

    tribution of SR in the medical field.

    In the rest of this introduction we introduce the non-medical

    reader to the variety of medical imaging modalities as well as

    introduce resolution-related challenges in the medical imaging

    field. A brief overview of the SR algorithms, SR parameteriza-

    tion, and evaluation schemes are given in Section 2. Section 3

    focuses on the application of SR in MRI, and Section 4

    describes major works on the application of SR in PET. Key

    issues and an overview of the outstanding challenges in this

    domain finalize this overview paper, in Section 5.

    1.1. Medical imaging acquisition process and clinical

    significance

    Figure 1 illustrates the general medical imagery acquisition

    process. Energy is acquired as it transverses the body (or

    organ) as part of a transmission or an emission process. For

    example, in X-ray computer tomography (CT), X-rays are

    transmitted through the body and captured by an array of

    detectors, following attenuation by the imaged object. In

    PET, photons are emitted from within the body (from radio-

    active molecules inserted into the body) and are then detected

    by an array of detectors. The energy is captured by an array of

    detectors that are designed per specific imaging modality [e.g.

    radio frequency (RF) coils in MRI, crystal detectors in PET].

    Figure 2 presents a schematic description of an X-ray trans-

    mission (a) and a photon emission process (b). Following

    the energy acquisition process, various algorithmic transform-

    ations are used, such as the inverse Radon transform (e.g. in

    CT), to reconstruct a visual image. The target of a successful

    acquisition process is a high spatial-resolution visual image of

    the organ of interest. Additional characteristics of interest

    include a high-contrast and a high dynamic-range as well as

    a strong signal-to-noise ratio (SNR) output from the imagingsystem.

    A complete physical and signal-processing review of the

    various medical imaging modalities is beyond the scope of

    this paper (several books can be found, including [710]).

    Below, we briefly introduce several key imaging modalities

    and present representative output images. Additional detail

    for each modality is given in Appendix A.

    MRI is an imaging technique used primarily in medical set-

    tings to produce high-quality images of the human body. It is

    based on the absorption and emission of energy in the

    RF range of the electromagnetic spectrum, producing images

    based on spatial variations in the frequency of the RF

    energy being absorbed and emitted by the imaged object.A sample MRI scanner and an MR reconstructed image of

    the human head are shown in Fig. 3a and b, respectively.

    A set of MRI brain scans, from three standard MR imaging

    sequences, termed T1, T2 and Proton-Density (Pd) (see

    Appendix A), are shown in Fig. 4. We note that each such

    sequence displays a unique image of the imaged organ

    (brain). A standard coordinate system was developed to rep-

    resent three (2-D) slice directions, as displayed in the figure.

    The defined slice directions will be used throughout the

    experiments reviewed in this paper.

    In addition to the standard, anatomical MR imaging, several

    variants exist, including the following: MRI Angiography

    deals with the imaging of the flowing blood in the arteries

    and veins of the body, with intensity proportional to the velo-

    city of the flow. MRI Angiography can be used to evaluate

    abnormal narrowing of the blood vessels (stenosis) and their

    risk of rapture (aneurysms). Diffusion-weighted imaging

    (DWI) is an MRI modality that produces the in vivo MR

    images of biological tissues weighted with the local character-

    istics of water diffusion. DWIs are very useful in diagnosing

    vascular strokes in the brain and to study white matter dis-

    eases. Diffusion tensor imaging (DTI) is a variation of DWI

    in which at least seven images are acquired for every slice,

    with at least six directions of diffusion weighting. DTI

    serves as a unique tool for visualization of the direction and

    intactness of white matter fiber tracts in vivo by identifying

    the preferred direction of diffusion. HR DTI combined with

    algorithms for tracing fibers in three dimensions in tensor

    fields has the potential to enable fiber tract mapping of critical

    functional pathways in the brain. The clinical applications are

    the tract-specific localization of white matter lesions, the

    localization of tumors relative to the white matter tracts

    and the localization of the main white matter tracts for neuro-

    surgical planning. fMRI measures the changes in blood flowFIGURE 1. Illustration of the general medical imagery acquisition

    process.

    44 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    3/21

    and blood oxygenation in the brain (hemodynamics), which iscorrelated to neural activity in the brain or spinal cord of

    humans or other animals. The magnetic resonance (MR)

    signal of blood is slightly different depending on the level of

    oxygenation. These differential signals can be detected using

    an appropriate MR pulse sequence.

    Figure 3 presents examples of an angiographic image (c), a

    DTI image (d) and an fMRI image (e).

    X-ray CT is based on the fact that X-rays can traverse a cross-

    section of an object along straight lines, be attenuated by the

    object, and detected outside it (Fig. 2a). Since its introduction

    in the 1970s, CT has become an important tool in medical

    imaging in the diagnosis of a large number of differentdisease entities. CT is currently a standard diagnosis tool in

    several domains, including: head imaging in particular for diag-

    nosis of cerebrovascular accidents and intracranial hemorrhage;

    facial and skull fractures evaluation; surgical planning for cra-

    niofacial and dentofacial deformities; detecting acute and

    chronic chest diseases; imaging of coronary arteries (cardiac

    CT angiography); abdominal diseases; imaging complex frac-

    tures especially around joints, and more. Recently, CT is also

    being considered for preventive medicine or screening for

    disease, for example CT colonography for patients with a

    high risk of colon cancer. The CT scanner can image complete

    organs and volumes using a large series of two-dimensional

    x-ray images taken around a single axis of rotation. Over

    recent years, a transition has been made from slice-by-slice

    imaging to volume imaging, with the introduction of spiral

    scan modes. An example CT image is shown in Fig. 6a.

    PET belongs to the radiology specialty of nuclear medicine,

    and it provides information on the distribution of a chosen

    molecule inside the human body (Fig. 2b). PET provides func-

    tional information that is complimentary to the anatomical

    information from other radiological imaging techniques such

    as the MRI and CT. In particular, PET is emerging as an

    important tool to detect tumors and to evaluate their degree

    of malignancy, based on differences in biochemistry and

    metabolism between tumors and their surrounding normal

    tissues [11]. PET is also used for functional brain imaging

    and in cardiology, to assess myocardial viability and efficiency

    [1215]. PET images are commonly fused with anatomical

    images such as CT. The need to combine PET and CT has

    evolved into specialized hardware that makes the task of

    fusing the two modalities much easier. A combined PET/CT

    machine is shown in Fig. 5. An example of PET image is

    shown in Fig. 6b and a combined PET/CT image is shown

    in Fig. 6c. For an overview on PET, see [16].

    FIGURE 2. Schematic description of x-ray transmission and PET acquisitions. (a) X-ray CT acquisition via X-ray transmission. (b) Schematics

    of PET acquisition via photon emission. The PET camera is comprised of a ring of discrete detectors. Shown on the left is a pair of opposing PET

    detectors. Each such detector pair detects coincident photon pair emission.

    FIGURE 3. The MRI modality. (a) MRI scanner (GE medical

    systems). (b) Anatomical image of a head. T1-weighted. (c) Angio-

    graphy. (d) DTI. 150 gradient directions. (e) fMRI. Activated areas

    overlayed on the anatomical image (for color see Figure 12).

    SR IN MEDICAL IMAGING 45

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    4/21

    FIGURE 4. Brain slices in varying MRI sequences (taken from BrainWeb [17]). Transverse slices are in the xy plane. Sagittal slices are in the zy

    plane. Coronal slices are in the zx plane.

    FIGURE 5. Combined CT/PET scanners. (a) Illustration. (b) Scanner (GE medical systems).

    FIGURE 6. Example brain scans. (a) CT image. (b) PET image. (c) PET and CT images fused together.

    46 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    5/21

    1.2. Resolution limitations and challenges in medical

    imaging

    A common goal in all medical imaging systems is to increase

    the resolution and, to the extent possible, achieve true isotro-

    pic 3-D imaging. To capture the entire frequency content of

    the imaged object, a sampling rate at the Nyquist frequency

    is required, defined as twice the highest frequency present inthe imaged object. Correspondingly, the sampling distance

    must be one-half the spatial resolution, defined as the distance

    between half-value points of the system impulse response, or

    the full-width-at-half-maximum (FWHM).

    In practice, the range of frequencies captured is limited by

    the maximal sampling frequency of the imaging device detec-

    tors, as defined by the detector pitch, or the detector spacing.

    Reduced size (width) of detectors and smaller inter-detector

    distances can provide increased resolution, but this is at the

    cost of an increase in the noise, thus results in a much-reduced

    SNR. The sensor resolution of a general imaging system is

    determined according to the physical constraints of the detec-

    tors, which are in turn tuned to SNR and timing considerations

    in the system. Within each imaging modality, specific physical

    laws are in control, defining the meaning of noise and the sen-

    sitivity of the imaging process. Signal processing rules govern

    the system design in an attempt to achieve an acceptable com-

    promise between resolution and SNR. The resolution limit-

    ations in MRI and PET will be defined in depth in Sections

    3 and 4, respectively.

    A resolution-related challenge in medical image processing

    is known as partial volume effect (PVE), which arises when an

    interface between two different tissues occur within a single

    voxel. The PVE is a direct consequence of limited resolution

    during the acquisition process. In general, PVE blurs theboundary between tissues and adds complexity to tissue

    characterizations. In MRI, the resulting image pixels display

    a gray level proportional to the weighted average of the

    signals stemming from neighboring tissues. The exact location

    of boundaries may be shifted thus introducing a major obstacle

    for anatomical MRI brain segmentation. In a CT image, each

    voxel represents the attenuation properties of a specific

    volume. When more than a single tissue is present within

    the voxel, the value will be some (nonlinear) average of the

    tissues attenuation properties. In DTI, where isotropic resolu-

    tion is particularly important, PVEs are a limiting factor in the

    analysis of directional and structural axonal connectivity.

    Increased resolution can help overcome or reduce the pro-blems associated with PVE.

    1.3. Can SR support the medical imaging challenges?

    SR reconstruction deals with combining several LR images to

    create a HR image. SR techniques have been suggested in

    recent years as a means for increasing resolution without alter-

    ing the existing imaging hardware. Thus, they can be seen as a

    means for extending current medical imaging resolution

    limitations. The goal of SR algorithms is to improve the

    image resolution in cases in which the image was under-

    sampled. Such cases involve the following: first, the imaged

    object has high-frequency content. Second, the sampling fre-

    quency as defined by the detectors, does not fulfill the

    Nyquist frequency; thus, aliasing and degradation in the high-frequency content can be observed. The SR process helps

    overcome the detector sampling limitations by practically

    increasing the sampling rate, and thus utilizing additional

    high-frequency information and reducing the aliasing

    effects. Note that in cases in which no frequencies higher

    than half of the detectors sampling frequency exist, SR will

    in effect result in the averaging of noise; in such cases, no

    additional improvements in the image resolution can be

    obtained by SR.

    2. SR ALGORITHMS: BRIEF OVERVIEW

    The term SR refers here to a technique in which several LR

    images, from different POV relative to the image object, are

    combined to obtain a higher-resolution image. Several SR

    reconstruction methodologies have been developed in the

    last two decades [18]. In initial works [e.g., 19], the frequency

    domain was used to demonstrate the ability to reconstruct one

    improved resolution image from several down-sampled noise-

    free versions of it, based on the spatial aliasing effect. The fre-

    quency domain approach was further generalized to noisy and

    blurred images in [20] and a spatial domain alternative was

    suggested in [21]. Further non-iterative spatial domain data

    fusion approaches were proposed in [22, 23].

    An iterative back-projection (IBP) method was proposed in

    [24]. This method starts with an initial guess of the outcome

    image, projects the initial result to simulate the LR measure-

    ments, and updates the temporary guess according to the simu-

    lation error. Further detail on the IBP method is provided in

    the following subsections.

    A set theoretic approach to SR was suggested in [25]. Here,

    convex sets are defined which represent tight constraints on

    the required image. Nonlinear constraints are combined

    within the restoration process and a projection onto convex

    sets (POCS) algorithm is utilized. A hybrid model that com-

    bines maximum-likelihood (ML) and POCS was suggested

    in [26]. More recent SR works aim at combining the SR

    approaches with regularization terms, e.g. in [27] fast and

    robust multi-frame SR is proposed using L1 norm minimiz-

    ation and robust regularization based on a bilateral prior to

    deal with different data and noise models.

    In this section, we mathematically define the general image

    acquisition procedure, and present the general formalism for

    SR. It is our goal to provide the reader with a brief overview

    of specific algorithms and related parameterization issues, as

    related to the research works reviewed herein.

    SR IN MEDICAL IMAGING 47

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    6/21

    2.1. Modeling the image acquisition process

    Given a set fykgk1N of LR images and assuming all LR images

    are degraded versions of the same original HR image, x, SR

    algorithms reconstruct a super-resolved HR image which by

    simulating the imaging process, results in images that best

    describe the LR measurements via some SR criterion.

    Mathematically, we can express each LR image yk as theresult of a sequence of operators on the original HR image

    source, x, consisting of geometrical warp, blurring and deci-

    mation, as in equation (1),

    yk fkxh # s vk; k f1; . . . ;Ng; 1

    where fk is the kth geometrical transformation of the image x to

    the same reference frame of acquisition for yk, h is a blur

    kerneloften referred to as the point spread function (PSF),

    defined by the properties of the lens and the imaging device,

    and vk is an additive noise. The symbol (*) is the convolution

    operator and #s represents the down sampling of an HR image

    to a LR grid by a factor s.

    2.2. The general formalism of SR

    We present next the general formalism for SR, as relevant both

    for the IBP approach of Irani and Peleg [24], as well as to the

    more recent, ML formalism of [27]. An initial estimate of the

    HR image, x(0), is taken as the average of the set of LR acqui-

    sitions brought to the same reference point and up-sampled:

    x0 1

    kX

    K

    k1

    f1k yk " s; 2

    where yk is one of K acquisitions, fk the geometric transform-

    ation to a common reference frame and "s the up-sampling

    operator from LR to the HR representation.

    Given an image-acquisition model, and an HR estimate (of the

    nth iteration), x(n), a set of synthetically generated LR images,

    fyk(n)g, can be extracted. The process involves shifting the HR

    image to the kth POV, blurring to account for the psf and down-

    sampling to the systems sampling rate. The nth LR synthetically

    sampled set of images fyk(n)g is thus obtained from the nth

    approximation of the HR image x(n) (ignoring the noise), by:

    ~ynk fkx

    nh # s: 3

    The current iteration x(n) is updated according to the differ-

    ence between the synthetically generated set of LR images

    fyk(n)g and the actual acquired set of LR images fykg:

    xn1 xn 1

    k

    XKk1

    f1k yk ~ynk " sp: 4

    The differences between the original image and the syn-

    thetically generated image are up-sampled to achieve the

    smaller SR pixel size, moved to a common reference frame,

    and averaged over K acquisitions. The symbol p is termed

    the BP kernel and is related to h [27] (often h is assumed

    to be symmetric which results in p h). The defined iterative

    process [equations (3) and (4)] is repeated until a predefinederror measure, such as the mean square error in the

    maximum-likelihood (ML) sense:

    12ML

    x XKk1

    yk ~ynk

    25

    has been minimized or has reached a predefined threshold. Alter-

    natively, a maximum number of iterations has been reached.

    Equation (1) defines a classical restoration problem for the

    original HR image x. The solution to this problem depends on

    the minimization criterion. Equations [3, 4] describe a solution

    in the sense of ML estimator that minimizes theL2

    normcriterion [equation (5)], which assumes an additive indepen-

    dent noise with normal distribution in the forward model.

    In [27], the solution for the general Lp norm is derived. It is

    shown that the L1 norm is more robust and can be used for

    strong sporadic noise.

    In SR reconstruction with MAP estimation, prior know-

    ledge on the imaged data, or the desired reconstructed

    solution, is incorporated as a regularization term in the mini-

    mization process:

    12MAP

    xn XKk1

    yk ~ynk

    2lA xn; 6

    where A(x) is the regularization term providing prior infor-

    mation on the desired SR image x, and l is the regularization

    coefficient specifying the weight of the regularization term.

    Several regularization terms from the literature have been

    used in SR, including the Tikhonov cost function [26] and

    the bilateral filter [27].

    2.3. SR parameters and performance evaluation:

    general and medical domain considerations

    Each SR algorithm has certain key parameters that need to be

    determined to most closely match with the true imaging

    system characteristics and specific application scenario. Two

    key parameters are the transformation and blur: the transform-

    ation parameter needs to enable precise image registration,

    accurate to a small fraction of a pixel, capable of bringing

    all input images to a common reference frame. In medical

    systems, the transformation is typically the physical shift

    (from an original position) between the object (patient bed)

    and the imager.

    48 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    7/21

    The blur, h, is defined by the imaging system PSF. In

    medical imaging systems, the PSF is often defined as the

    slice excitation profile. Experiments indicate that the typical

    slice profiles are well approximated by Gaussian functions,

    where the FWHM is the originally selected slice width. Two

    PSFs are thus commonly used to represent the blur: the first

    is a rectangular pulse PSF which is a crude estimation forthe slice profile (termed Box-PSF), where the box width is

    taken as the selected slice width (in the desired HR pixel

    units). The more accurate representation is the use of a Gaus-

    sian PSF (Gaussian-PSF), with FWHM set to the selected

    slice width.

    The BP parameter of equation (4), p, is ideally the inverse of

    the blur kernel. In [24], it was proven that the proposed SR

    algorithm converges with a convergence condition (for 2-D

    translations and rotations) given by kd2 h *pk, 1, where

    d is the unity impulse function. The smaller kd2 h * pk is

    the faster the algorithm converges. This criterion allows the

    kernel p to be other than the exact inverse of the blurring func-

    tion h. For the typical slice profiles mentioned above, the pfilter can be taken as an impulse function, satisfying the

    Irani Peleg requirement for convergence. In both MRI [3]

    and PET studies [5], the blur h and BP kernel p were corre-

    spondingly set to unity. In [6], an increase in model accuracy

    is attempted by defining both the blur kernel as well as the BP

    kernel, to be modeled as a Gaussian PSF.

    In regularization-based schemes, an additional minimiz-

    ation parameter exists per defined objective function. Optimi-

    zation of this parameter can be performed in all three spatial

    directions, denoted as 3-D anisotropic filtering, or in the

    slice select direction only, known as 1-D anisotropic

    filtering.

    Quantitative measures for any image enhancement pro-

    cedure are a challenge. In general, image enhancement can

    be expressed as the increase in edge slope in the image

    plane as well as the increase in frequency content as viewed

    via the image power spectrum. When considering a method

    for resolution improvement, it is important to ensure that the

    SNR is not compromised. The SNR is measured by taking

    the mean of a high-intensity region of interest and dividing

    by the standard deviation of a region of noise outside the

    imaged object. A variation of the above measure is a contrast

    ratio measure, C, which defines the ratio between the average

    signals to the average background.

    Traditional alternatives to SR reconstruction include zero-

    padding, or sinc interpolation, and interleaving. In zero-

    padding, the spectral resolution is augmented by adding

    zeros embedded between the given samples. Padding the

    data with zeroes provides more frequency-domain points

    (improved spectral resolution), but does not improve the resol-

    ution limits as established by the given sampling rate, nor does

    it alter the effects of aliasing error. Interleaving is a method for

    achieving an HR image from a set of shifted LR images by

    combining the pixels, one by one, from alternating LR

    image inputs, to generate a single large image. In the studies

    reviewed below, the two alternatives are utilized for compari-

    son and evaluation of the SR methodology in varying medical

    imaging modalities.

    3. SR IN MRI

    In this section, we focus on the application of SR method-

    ologies to the MR imaging modality. We start by highlighting

    the key factors that limit resolution in MRI. We then present

    an analysis into what can be termed the SR dimensionality

    in MRI. An extended overview of recent works that explore

    the possibility of augmenting MRI resolution, using SR algori-

    thms, is presented next. Both anatomical as well as functional

    MR images have been used with promising initial results. In

    [3], the IBP SR method was used for anatomical MRI.

    Results were shown on both phantom as well as human

    brain data and demonstrated that isotropic resolution can be

    achieved while preserving SNR. SR reconstruction based onthe use of discontinuity-preserving regularization methods

    was proposed in [4] for HR fMRI image reconstruction.

    Results demonstrate that the use of SR may increase the

    ability to detect and visualize small regions of neuronal

    activity; moreover, the activated regions appear sharper and

    provide better information regarding their morphological

    limits and structure.

    3.1. Resolution challenges in MRI

    High resolution, isotropic 3-D MRI images are important for

    visualization of 3-D volumes in the imaged object and for

    early medical diagnosis. In practice, true 3-D acquisition

    methods are frequently not effective or possible, as is often

    the case in T2-weighted imaging, DWI and occasionally in

    MR angiography (MRA). True T2-weighting is difficult to

    obtain in reasonable imaging times by 3-D acquisition

    methods. The problem arises due to the need for long signal

    recovery between excitations to enable the operation of

    the spin-echo mechanism that provides T2 contrast (see

    Appendix A). Since all the spins are excited by every pulse,

    the recovery time cannot be utilized and the sequence takes

    a long time. In DWI, no 3-D technique for humans currently

    exists. Sequences that acquire raw data pertaining to the

    same slice or volume over many excitations cannot be modi-

    fied to provide diffusion-weighted contrast because of phase

    inconsistencies resulting from physiological motion. MRA is

    another popular application that sometimes performs better

    in the 2-D rather than the 3-D version. In fMRI, temporal res-

    olution as well as spatial resolution is important. Most 3-D

    acquisition procedures cannot reach the required temporal res-

    olution necessary for appropriate statistical analysis.

    When the true 3-D image acquisition is not effective or

    possible, it is common practice to acquire a set of 2-D

    SR IN MEDICAL IMAGING 49

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    8/21

    slices. The problem, as illustrated in Fig. 7, is that a set of 2-D

    slices does not give a good isotropic 3-D image. A recon-

    structed MR image is commonly of HR in-plane (x, y) and

    of much reduced resolution in the slice-select (z) direction.

    For example, it is common to find reconstructed MR images

    of size 1 1 3 mm3. The spatial resolution in-plane (x, y)

    is determined by several factors, including the gradients

    intensity, the imaging bandwidth, the number of readout

    points and phase encoding steps (see Appendix A and

    related references). The slice thickness in MRI is determined

    by what is termed the slice-selection pulse, which is in turn

    determined by hardware limitations coupled with pulse

    sequence timing considerations. The challenge for SR in

    MRI is to increase the resolution in the slice-select dimension

    so as to achieve HR, isotropic, 3-D images. A further chal-

    lenge is to achieve the HR outcome without decreasing the

    SNR.

    3.2. SR dimensionality

    Several works have used theoretical analysis and experimental

    validation, to reveal the potential for SR in MRI along with the

    appropriate task dimensionality [3, 28]. Signal-processing

    principles underlying MRI acquisition, in particular, in

    Fourier-encoded MRI data sets, have led researchers to

    acknowledge the distinct characteristics of the in-plane

    versus slice-select encoding. This fact, in turn, affects the

    effective dimensionality of the SR task. In particular, Fourier-

    encoded in-plane MRI data is inherently band limited. This is

    due to the time limit of the acquisition process and the fact that

    the information is gathered in the frequency domain (known as

    k-space acquisition). In-plane shifting is thus equivalent to a

    global phase shift in the acquisition space (k-space), the orig-

    inal temporal domain, which does not affect the inherent

    spatial frequency resolution of the acquired data. In other

    words, increasing the in-plane resolution by in-plane shifting

    of the image is equivalent to zero-padding of the raw data in

    the temporal domain. A different scenario exists in the

    slice-select direction of a Fourier-encoded MRI. There is

    sufficient information in the slice-select dimension such that

    under-sampling of the data in that direction results in aliasing.

    A less sharp cut-off can thus be observed when viewing thespatial frequencies in the slice-select (z) direction in Fourier-

    encoded MRI. The existing aliasing in the slice-select direc-

    tion provides the basis for using SR algorithms in enhancing

    the resolution.

    The illustration shown in Fig. 8 (taken from [3]) is used to

    validate the above claims: multislice 2-D image data sets were

    acquired, with half-voxel shifts in all three spatial directions.

    A 3-D iterative SR algorithm was applied [3]. The original

    LR image is shown (top left) with its original power spectrum

    (top right). The output of the SR process (double size in each

    dimension) is shown (bottom left) with its power spectrum

    (bottom right). The sharp frequency cut-off in the y

    (in-plane) direction is evident. A spreading out of the power-spectrum is present in the slice-select (z) direction, indicating

    an effective augmentation in the resolution. The conclusion

    regarding the dimensionality of the SR task is the following:

    the best that can be done in the in-plane (x, y) is to interpolate,

    via zero-padding, the given data to the desired resolution. In

    the slice-select dimension, sub-voxel spatial shifts can in

    fact be used to increase the resolution. In current Fourier-

    encoded MRI systems, the task is inherently a 1-D, slice-select

    task. An interesting point is that if successful in this dimen-

    sion, the goal of 3-D isotropy can in fact be achieved.

    3.3. Experiments and results: anatomical MRI

    In [3], the possibility of using SR for inter-slice MRI data was

    explored. Several key results are reviewed herein. The SR

    algorithm used was the Irani Peleg IBP method of [24]

    (Section 2). Anatomical MRI results are shown on a

    phantom, on inanimate objects and on a human brain. Quali-

    tative results are shown, followed by quantitative evaluation.

    All imaging was performed with an RF head coil on either a

    1.5 Tesla GE Signa MRI system or a 3 Tesla GE MRI

    FIGURE 7. MRI slice acquisition. The resolution in the slice-select

    direction is lower than in the in-plane directions [3]. FIGURE 8. Investigating SR dimensionality in MRI. Spectrum

    analysis (y,z) plane, Horizontal axis is the slice-select (z) direction.

    Top row: LR input (left) with spectrum (right). Bottom row: HR

    output (left) and spectrum (right) [3].

    50 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    9/21

    system. The phantom used for the experiment consisted of

    long thin plastic partitions (teeth), lodged in a plastic

    block, placed 4 mm apart, surrounded by Gd-DTPA-doped

    water. The imaging sequence consisted of multislice fast

    spin-echo (FSE) with 16 slices, 3-mm thick, approximately

    parallel to the plastic partitions. Three sets of multislice data

    were acquired, with 1 mm shifts in the slice-select direction.

    The LR input voxel size was 1 1 3 mm3. Following the

    SR procedure, an output voxel is a 1 mm isotropic cube.

    Results are shown in Fig. 9. The visibility of the comb teeth

    has greatly improved by using SR rather than zero-padding

    interpolation. Moreover, more information is evident with

    SR than with interleaving. The implementation with a

    GaussianPSF (e) seems to give slightly better results than

    when using a box-PSF (d): Better estimation of the HR

    image is achieved by using a blurring filter, h, that more

    closely matches with the MRI system and the MR image

    characteristics in the slice-select dimension.

    Results on an inanimate object are shown next. A papaya

    image (zoomed-in) is shown in Fig. 10. The x-axis is the slice-

    select axis. The input LR image, shown in Fig. 10a, is of res-

    olution 1 1 3 mm3, whereas the image following SR

    shown in Fig. 10b has a resolution of 1 1 1 mm3. To

    quantify the resolution augmentation, both the image and fre-

    quency domains are used. Figure 10c and d shows a

    comparison of edges (two examples) within the two images,

    while Fig. 10e and f shows the augmentation of the frequency

    spectrum in the slice-select direction.

    Figure 11 shows initial SR results on human brain data. An

    FSE imaging sequence was used with three shifts in the slice

    direction. The original slice thickness was 4.5 mm, in-plane

    resolution was 1.5 mm and the number of slices was 22. SRoutput resolution was 1.5 mm (cubed). Results show a clear

    improvement in the progression from the LR input

    (Fig. 11a) or the zero-padded input (Fig. 11b) to the SR

    results (Fig. 11c and d).

    Quantitative evaluation was carried out on an apple input

    source (in [3] and revalidated in [4]). The resolution was quan-

    tified by the measurement of edge widths (see [3] for details on

    measuring width). Timing required and measured SNR were

    recorded as well. Table 1 summarizes the main results.

    Looking across the rows it is clear that the resolution in edge-

    width improves as a shift is made from the zero-padded input

    to the HR source. The SR result is much better than the zero-

    padded or interleaved result, with the mean edge width of theSR result almost identical to the mean width of the HR source.

    SNR values decrease with the increased resolution. Note that

    the SNR of the SR result is higher than the SNR of 2-D thin

    slice acquisitions. In an MRI process, the goal is to obtain

    HR images with a high-SNR efficiency, which is the ratio

    between the SNR of the result and the square root of the

    time length of the data acquisition sequence. The SNR effi-

    ciency measure displays a similar trend: an SNR efficiency

    of 7.63 is achieved for the SR result, compared with 6.13 for

    the HR acquisition.

    3.4. Experiments and results: fMRI

    Peeters et al. [4] propose an optimization approach for HR

    fMRI reconstruction using SR. The fMRI acquisition is

    adapted to acquire two image stacks with low slice resolution,

    shifted over half-a-slice thickness. Two separate slice-shifted

    overlapping volumes are acquired, each obtained at half the

    acquisition time of the HR volume. The shifted volumes are

    combined via SR to reconstruct a stack of slices with half

    the acquisition thickness. The SR methodology used in this

    work is based on edge-preserving approaches and conver-

    gence rate studies [29].

    fMRI data differs from anatomical MRI data in that it

    involves dynamical data; it images the hemodynamic response

    function (hrf). It is important to note that the image data

    acquired during the initial and final portion of the hrf is not

    in a stationary state (plateau), and thus cannot be utilized in

    SR algorithms, which assume a combination of two sets of

    volumes acquired with an identical neurophysiological

    response condition. In [4], the solution suggested to this

    dynamics issue is the elimination of specific volumes that

    were acquired during the non-stationary state of the response.

    FIGURE 9. SR on a phantom image. (a) The original LR data. (b)

    Zero-padding interpolation. (c) Interleaving the slices. (d) SR with

    box-PSF. (e) SR with Gaussian-PSF. Horizontal axis is the slice-

    select axis [3].

    SR IN MEDICAL IMAGING 51

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    10/21

    Results of using SR on fMRI datasets, both simulated and real

    data are shown next, based on [4].

    In the real fMRI datasets, a visual stimulation paradigm for

    retinotopic mapping was used. The stimuli used were designed

    to stimulate the horizontal (HM) and vertical (VM) visual field

    meridian, using horizontally and vertically oriented wedge-

    shaped checkerboards alternating at 4 Hz. The HM and VM

    stimuli were alternated in blocks of 10 brain volume scans.

    In this experiment, a total of 10 sessions of 12 blocks each,

    i.e. 120 scans per session, were performed on the same

    subject. During the experiment two different acquisition strat-

    egies were interleaved: the HR fMRI (ground truth) acqui-

    sition protocol and the LR (slice shifted) fMRI acquisition

    protocol, yielding five high resolution (ground truth) and

    FIGURE 10. Papaya example. (a) LR data. (b) SR result. (c) and (d) Comparison between two corresponding edges of the input image (dotted line)

    and the SR image (solid line). (e) Power spectrum of input image. (f) Power spectrum of SR image.

    FIGURE 11. Human brain MRI. (a) The original LR data. (b) Zero-padding interpolation. (c) SR with box-PSF. (d) SR with Gaussian-PSF [3].

    52 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    11/21

    five slice-shifted LR volume datasets. In the slice-shifted

    mode, the slice thickness was doubled as compared to the stan-

    dard sequence. The two shifted volumes were acquired con-secutively, with the second volume shifted in a slice position

    over a distance equal to the slice thickness of the HR

    images. The high-resolution data were collected on a

    Siemens Sonata 1.5 Tesla MR system. A matrix of 128

    128 was acquired. Voxel size was taken as 2 2 2 mm3

    for the ground-truth (HR) images and 2 2 4 mm3 for the

    slice-shifted images. The global acquisition time of the high-

    resolution volume was 3328 ms and each of the slice-shifted

    interleaved volumes was 1664 ms (for additional acquisition

    protocol details see [4]).

    Figure 12 shows statistical parametric mapping (SPM) acti-

    vation maps [30], displaying the activated areas above a stati-

    stical threshold (pcorr, 0.05), overlaid on the mean EPI slices

    of the interpolated datasets. Patches of color on the MRI brain

    slice shows differences in brain activity, with the colors repre-

    senting the location of voxels that have shown statistically sig-

    nificant differences between experimental conditions (see

    Appendix A). Figure 12a displays the images of the retino-

    topic mapping fMRI experiment in the acquisition plane

    with the corresponding activation superposed, and Fig. 12b

    shows activation images perpendicular to the slice direction.

    Cases compared include the following (top to bottom, left to

    right): original HR and LR inputs, an average dataset in

    which each HR voxel is computed as the average of all LR

    voxels that contain it (equivalent to the initial guess step of

    Irani Peleg), two SR results, and a composed result,

    which is equivalent to interleaving. In the SR algorithm, the

    regularization term was applied in all three dimensions (Ani-

    sotropic 3-D) or in the slice-select only (Anisotropic 1-D).

    Qualitative analysis of the results indicates overall simi-

    larity between the interpolated datasets and the reference

    HR dataset. A closer look at the data reveals that the 1-D

    and 3-D anisotropic SR datasets show higher t values at the

    foci of the activated areas than the reference HR, the original

    LR and the average and composed data sets. Also the size of

    the activated clusters with a t value above threshold appears

    to be larger in the SR datasets compared to the original data.

    When comparing the two different SR datasets (3-D versus

    1-D anisotropic regularization), the following differences are

    observed: the 3-D anisotropic images display more intense

    but less sharp activation patches than the 1-D anisotropicinterpolated and HR dataset both in- and through-plane. The

    1-D anisotropic SR dataset demonstrates a higher resolution

    of the activated patches in the slice direction than the LR

    dataset and the 3-D SR anisotropic dataset. These results can

    be explained via the larger smoothing inherent to the 3-D ani-

    sotropic interpolation algorithm in all directions.

    A second synthetic database was generated with known acti-

    vation areas inserted, as shown in Fig. 13. The mean EPI MR

    volume was used as a template with a base resolution of 3

    3 4 mm3. This template was duplicated 120 times to gener-

    ate a dynamic time series, with different activated regions

    inserted in an interleaved mode of 10 rest volumes and 10

    activated volumes. These activation regions consisted ofdifferent spheres with different radii and an irregularly

    shaped area at carefully chosen positions. The intensity of acti-

    vation was set to a maximum of 8% peak signal change. Two

    slice-shifted LR datasets were generated by addition of the

    adjacent slices of the HR dataset. Gaussian noise was inserted

    with a standard deviation of 2% for the HR set and 1% for the

    LR volumes. Figure 13 shows the statistical analysis on

    the synthetic dataset, with similar conclusions obtained as in

    the real dataset. Both 1-D and 3-D SR anisotropic datasets

    extract the activated areas in good agreement with the position

    and size of the simulated activated areas.

    Quantitative measures of the ability to separate two

    close-by-activated areas were extracted. Figure 14 shows the

    line graphs of calculated t-values of a cut in the slice direction

    through two activated regions, with a separation of two slices

    in Fig. 14a and a single slice in Fig. 14b. The results demon-

    strate that the SR algorithms show a good separation, closely

    resembling the HR dataset and much better that in the LR

    input and average or interleaved datasets.

    4. SR IN PET

    In this section, we focus on the application of SR to PET. We

    discuss the resolution limitations in PET and the challenges for

    SR in PET. Recent works that have started to investigate the

    potential of SR in this domain are reviewed herein.

    PET resolution is limited by physical properties, such as

    scatter, counting statistics, positron range and patient

    motion, as well as by the detector array geometry and the

    implemented acquisition protocol. Detector widths are

    limited to a certain minimal size, due to SNR considerations.

    A width that is too small will reduce detection efficiency and

    will increase intercrystal scatter and penetration. Resolution in

    TABLE 1. Quantitative measures of SNR and resolution on apple

    FSE sequence

    MRI on an apple

    input source

    Acquisition

    time (min:s)

    Mean

    edge width

    (pixels)

    SNR SNR

    efficiency

    (s21/2

    )

    Zero-padded

    reconstruction

    1:28 3.7 287 30.6

    Interleaved

    reconstruction

    4:24 3.7 276 16.97

    SR reconstruction

    box PSF

    4:24 2.9 170 10.46

    SR reconstruction

    Gaussian PSF

    4:24 2.2 124 7.63

    High resolution 4:00 2.3 95 6.13

    Original slice width of 4.5 mm; three shifts.

    SR IN MEDICAL IMAGING 53

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    12/21

    PET scanners is often degraded in order to achieve an accep-

    table image variance, where the variance is largely determined

    by the number of counts (counting statistics) collected during a

    scan. The noise that affects the counting statistics is comprised

    of several factors, as illustrated in Fig. 15. The first one is the

    angular uncertainty of the photons created in the annihilation

    process. Although the photons emitted in this process should

    move in a straight line of 1808 with respect to each other,

    there is a small angular divergence. The second limiting

    factor is the scatter events that one or both of the photons

    may pass before they reach the detector. The scatter event

    causes miss-estimation of the line where the annihilation

    process took place. Also present are random events that

    occur simultaneously and introduce wrong information to

    the reconstructed image.

    Current clinical PET scanners consist of 18 39 rings of

    detectors, which are aligned axially. A volumetric PET data

    set is commonly reconstructed by collecting a stack of 2-D

    transaxial images perpendicular to the axial (bed) direction.

    Many PET scanners have the option of restricting the

    line-of-response for gamma-ray coincident pair detection to

    the transverse plane perpendicular to the axial direction.

    This restrictive acquisition mode is termed the 2-D acquisition

    mode. A 3-D acquisition mode is one in which no restrictions

    apply during the acquisition process thus maximizing the

    number of events detected. In this scenario, the data are typi-

    cally rebinned into transverse planes [32] and then recon-

    structed using a 2-D algorithm that generates images from

    projection data [33]. In either mode, a reconstructed 3-D

    PET data set consists of a stack of 2-D transverse images

    FIGURE 12. Slices of activated areas resulting from the visual stimulation paradigm overlaid on mean EPI images for different datasets.

    (a) Transversal acquisition plane. (b) Sagittal slices [4].

    54 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    13/21

    along an axial direction. Coronal or sagittal images are gene-

    rated by re-sampling the voxel matrix along these planes.

    Accordingly, the spatial resolution in the transaxial plane is

    largely limited by the detector width, whereas the resolution

    along the axial direction is affected by the spacing of the

    detector rings. In practice, the final reconstructed resolution

    of a PET image is usually poorer than the best obtainable,

    intrinsic resolution, because reconstruction algorithms typi-

    cally trade-off resolution for reduced noise. In [16], an

    example is given where the intrinsic resolution is ,5 mm

    yet the final resolution of the image is greater than 8 mm.

    A typical resolution in clinical scanners is between 4 and

    7 mm FWHM [31].

    The SNR considerations, as briefly outlined above result in

    an under-sampling of available data. Wider detectors define a

    lower sampling frequency, while preserving a high-SNR ratio.

    This trade-off ensures that PET is a good candidate for SR.

    Higher resolution PET images may have several implications

    in research and clinical practice. The imaging of small cer-

    ebral structures such as the cortical sub-layers and nuclei

    may need PET spatial resolutions of ,2 mm [34, 35].

    Higher PET resolution would also be beneficial for improving

    FIGURE 13. Comparison between activated areas observed in the high, low and interpolated datasets for the synthetic data. (a) Transversal

    plane. (b) Sagittal slices [4].

    FIGURE 14. Line graphs of a cut in the slice direction showing the z-score for activated areas separated by two slices (a) and one slice (b) for

    different reconstructions. Real boundaries of the activated areas are shown with a solid black line (for color see [4]).

    SR IN MEDICAL IMAGING 55

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    14/21

    sensitivity for detection of small tumors [36]. Cancer lesions

    need to be of diameters equal or larger than the resolution of

    the PET scanner to be identified provided they also have ahigh-glucose metabolism. Finally, higher resolution PET

    images may show a more differentiated anatomical structure.

    The increase in anatomical detail may aid in the registration of

    a PET image with a corresponding anatomical image from

    another modality, such as CT or MRI.

    In recent work, the possibility of augmenting PET resol-

    ution using SR algorithms was explored [5, 6]. In [5], the

    use of SR to improve PET resolution using shifts and rotations

    in the transaxial plane as well as along the axial direction was

    demonstrated (directions illustrated in Fig. 15). Motivated by

    the results of SR in MRI [3], the IraniPeleg SR algorithm

    was used, with results demonstrated on a phantom as well as

    initial patient data. In a phantom study, the SR technique

    was shown to improve resolution and increase the contrast

    ratio, using a commercially available PET scanner, without

    increasing total scan time. In the patient study, an increase

    in scan time for one field of view (FOV) demonstrated that

    it is feasible to apply SR axially in a clinical setting without

    increasing the radiation dosage and without the need for any

    modification to the PET scanner hardware.

    4.1. Experiments and results

    In [5], an SR scheme based on the Irani Peleg iterative algor-

    ithm is proposed and experimentally confirmed to improve the

    resolution of PET. SR attenuation corrected PET scans of a

    phantom were obtained using the 2-D and 3-D acquisition

    modes of a clinical PET/CT scanner (Discovery-LS PET/CT

    scanner, GE Medical Systems). A special phantom was con-

    structed as shown in Fig. 16. The phantom contained holes

    of sizes 1, 1.5, 2, 4, 6 and 8 mm in diameter. Several exper-

    imental scenarios were tested: 1-D axial shifts (equivalent to

    bed-shifts), combined axial and transaxial shifts, and a

    patient study for the detection of small lung lesions.

    In the 1-D mode, SR via Irani and Peleg [24] was

    implemented along the axial direction by combining the sets

    of four LR acquisitions, each shifted by one-fourth of a LR

    pixel relative to the previous one [similar method used in [3]

    for the anatomical MRI (Section 3)]. In a combination of1-D (axial) and 2-D (transaxial) shifts, otherwise termed the

    3-D brain acquisition mode, SR was applied to transverse

    images using rotation and translation in the transverse plane

    as the geometric shift between successive acquisitions.

    Small CT markers were attached to the phantom in order to

    track the shifts.

    In the patient study, CT images were evaluated to identify

    regions-of-interest exhibiting the suspicious small lung

    lesions. Following a PET/CT scan, the patient was requested

    to remain still and bed was positioned so that the ROI was cen-

    tered in the FOV of the PET scanner. The scanner was pro-

    grammed for four additional PET acquisitions of this FOV,

    lasting 4 min each. Between each acquisition, the bed wasautomatically shifted by 1 mm. The patient was not exposed

    to any additional radiation since the X-ray CT component of

    the PET/CT study was not repeated. Both standard and SR

    images were processed, with additional re-slicing to provide

    coronal, sagittal and transverse images through the lesion of

    interest. In all the scenarios above, the process of synthetically

    generating LR acquisitions, comparing them to the four

    measured acquisitions, and updating the HR estimate was

    repeated until a predefined error was reached, or 16 iterations.

    The final estimated HR image was referred to as the SR result.

    Figure 17 shows the 3-D brain acquisition mode in which

    both axial as well as transaxial shifts are conducted (combined

    1-D and 2-D). CT markers provided the data needed to deter-

    mine the geometric shifts between successive images. In this

    example, between the initial PET image (Fig. 17b) and the

    FIGURE 15. PET events. (a) Graphic representation of true.

    (b) Scatter. (c) Random events [31].

    FIGURE 16. Phantom disk for PET experimentations ([5], #2006

    IEEE).

    56 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    15/21

    fourth PET image (Fig. 17c), there is a rotation of 7.28 and a

    translation of 9.4 mm. Reconstructed images are shown in

    Fig. 18. The top two images are coronal images and the

    bottom two images are transaxial images. On the left are

    images generated using a standard PET acquisition and on

    the right are SR images. All images were taken with the

    same total acquisition time. Both 4 (gray arrow) and 3 mm

    (black arrow) holes are more distinct in the SR images than

    in the standard acquisition images, for both coronary as well

    as the transaxial cases. Table 2 indicates that the SR image

    consistently provides a better contrast ratio than the other

    methods (see definition of Contrast in Section 2).

    An additional computational measure is the PSF FWHM,

    which can be computed from an approximate point source.

    Table 3 shows the case of a 1-mm diameter circular hole.

    The axial resolution in the 2-D whole-body mode was calcu-

    lated to be 4.1 mm FWHM with the SR algorithm, which is

    superior to both interleaving (4.9 mm FWHM) and the stan-

    dard reconstruction (4.8 8.6 mm FWHM). Thus, it can be

    concluded, that for the 3-D brain scenario, SR provides

    better resolution than the other methods both axially andtransaxially.

    Figure 19 shows the SR data in a study of a patient with a

    suspicious small lung lesion on CT. The uptake in the small

    lesion seen in the SR image of Fig. 19c is more localized

    than that seen on the corresponding standard original PET

    image in Fig. 19b. The second uptake (gray arrow) does not

    appear clearly in the SR image of Fig. 19c; rather it falls in

    an adjacent image plane, shown in Fig. 19d.

    FIGURE 17. 3-D brain mode acquisition. (a) CT with transaxial

    rotation and translation, four acquisitions, gray arrow indicates one

    of the CT markers. (b) Initial transaxial PET image. (c) Fourth PET

    image ([5],#2006 IEEE).

    TABLE 2. Contrast ratio C for the PET signals in 3-D AC brain

    mode acquisition phantom trials ([5] #2006 IEEE)

    Image type 3-mm

    holes

    4-mm

    holes

    6-mm

    holes

    8-mm

    holes

    Coronal

    No SR N/R

    a

    N/R 2.1 2.9Four acquisitions

    interleaved

    1.1 1.3 2.4 3.0

    SR 1.1 1.5 2.8 3.5

    Transaxial

    No SR N/R 1.2 2.2 2.7

    SR 1.2 1.8 3.2 3.8

    aN/R, not resolved.

    TABLE 3. PSF FWHM values for phantom trials ([5]#2006 IEEE)

    Acquisitionmode

    Axis Standard(mm)

    Interleave(mm)

    SR(mm)

    2-D whole body Axial 4.8 to 8.6a

    4.1

    3-D brain Axial 5.3 to 8.7 4.8

    3-D brain Radial 5.2 to 5.5 N/Ab

    4.3

    3-D brain Tangential 4.9 to 5.2 N/A 4.3

    aPoint source either centered in a pixel (lower value) or between

    two pixels (upper value).bN/A, not applicable.

    FIGURE 18. 3-D brain mode acquisitionresults. (a) Standard

    coronal PET image. (b) Coronal image with SR. (c) Standard transax-

    ial PET image. (d) Transaxial with SR ([5], #2006 IEEE).

    SR IN MEDICAL IMAGING 57

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    16/21

    5. SUMMARY AND CHALLENGES AHEAD

    Recent studies using SR in medical applications have demon-

    strated that using SR technology enables the limits on slice

    thickness as posed by the physical properties of existing

    imaging hardware to be effectively broken. Higher resolution

    in the image plane usually means acquisition with a smaller

    sampling distance, by using a smaller detector (e.g. in PET)

    or by using higher magnetic field scanners and shorter

    sampling distances (in MRI). Due to physical constraints,

    HR image acquisition results in a lower SNR, i.e. a trade-off

    exists between resolution and SNR. One of the key features

    in SR technology is the ability to obtain an HR image with

    almost the same SNR as the original LR images from which

    it is constructed. In the current overview, we have demon-

    strated this fact for MRI, fMRI and PET.

    In both MRI and fMRI, reconstructed SR images displayed

    a close resemblance to the HR data, while improving the SNR.

    Two different SR algorithms were used in the reviewed study:

    the Irani Peleg iterative BP algorithm [2, 3, 5] and a

    minimization algorithm with a constraint term on the smooth-

    ness of the solution [4]. Overall, the two approaches display a

    similar set of results and conclusions. In the anisotropic SR

    technique of [4], an investigation was conducted into theeffect of 1-D versus 3-D smoothing. Quantitative comparison

    of the activation maps indicates that the 3-D anisotropic diffu-

    sion SR data set provides the largest response and the largest

    activated areas, with the extracted regions much smoother

    than the 1-D case. Thus, for increasing the detection capability

    of small-activated areas, a 1-D smoothing filter is to be chosen.

    In the PET phantom study [5], smaller features were

    resolved with SR than without (3-mm features, as opposed

    to the minimum of 4-mm in standard techniques), furthermore

    the features that were resolved have a higher SNR. The

    phantom trials showed improvement in both the axial and

    transaxial resolutions. The axial resolution was improved by

    952% compared to the standard method and by 1416%

    compared to the interleaving reconstruction method. In the

    3-D brain mode transaxial images, SR improved the resolution

    by 12%. In the patient study, SR displayed more accurate

    (more localized) 18F-FDG uptake, without using any hardware

    changes or any increase in the patient radiation exposure. In a

    recently published study [6], two main extensions were inves-

    tigated: first, a more accurate SR algorithm was defined, in

    which both the blur kernel as well as the BP kernel are

    FIGURE 19. Coronal (left), sagittal (middle) and transaxial (right) sections of one FOV of a patient. The white arrows and black arrows denote

    the lesion of interest. (a) X-ray CT scan. (b) Original non-AC18

    FFDG PET image using clinical reconstruction protocols. Image planes dis-

    played are 4-mm thick. (c) Non-AC PET image through the center of the lesion of interest using SR. (d) Non-AC PET image using SR. The sec-

    ondary foci of uptake (gray arrows) seen in the original images (b) are evident here, in super-resolution images of planes adjacent to those depicted

    in (c) ([5], #2006 IEEE).

    58 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    17/21

    modeled with a Gaussian PSF. This, more accurate modeling,

    improves the above phantom results by an additional 24%.

    A second contribution of [6] is the fusion of PET and CT

    data: in addition to augmenting the PET resolution via SR, a

    post-processing step combines smoothing and edge-

    enhancement of the resultant image, based on the HR border

    information from the CT. The combined SR and anatomicaledge information were evaluated in phantom and patient

    studies using a clinical PET scanner. In both cases, a substan-

    tial increase in contrast was demonstrated.

    SR, for images and image sequences, has customarily been

    treated as a 2-D problem. In the overviewed studies, SR has

    often been applied to 1-D signals. In the general medical

    arena, the extension of the SR concept to 3-D is strongly motiv-

    ated, as has been recently proposed in [3]. Practical constraints

    have so far limited actual usage of true 3-D SR algorithms. In

    Fourier-encoded MRI, in-plane resolution is constrained

    (Section 3). The 3-D problem is thus downgraded to a 1-D

    task. It may be the case that with the new MRI technology

    that is not Fourier-based, the possibility for 3-D SR mayarise. The PET SR example demonstrates SR in more than a

    single dimension. It seems that true 3-D may be applicable in

    this domain and may be a worthwhile effort for the future.

    Regardless of the dimensionality of the task, an important con-

    tribution of SR is the reduction of PVEs in the reconstructed

    image. In this respect, even if a SR algorithm is applied in a

    single (axial) dimension, it in effect contributes to the increased

    resolution in additional (transaxial) dimensions as well.

    5.1. Spiral CT: a case where SR does not work?

    An important question to address is the applicability of SR to a

    given medical imaging system. It is important to note that SR

    can augment the resolution as acquired by the system

    detectors, in cases in which the detectors have under-sampled

    the input data. In other words, high frequencies exist in the

    signal that reaches the detectors, and the detectors sampling

    limit leads to aliasing and degradation in the high-spatial fre-

    quency content, as output in the reconstructed image. SR

    reconstructs the aliased high frequency information thus pro-

    viding a higher resolution output and minimizing the aliasing

    problem. In cases in which no frequencies exist that are higher

    than half of the detectors sampling frequency, no additional

    improvements in the image resolution can be obtained by

    SR technology.

    Such is the case in recently developed spiral CT systems.

    Today it is possible to scan the complete body trunk with sub-

    millimeter isotropic resolution in less than 30 seconds, with an

    effective slice thickness of 0.10.2 mm. Although the detec-

    tors sampling looks very promising, it is difficult to achieve

    this resolution in the reconstructed image. The main reason

    is the PSF of the spiral CT acquisition system, which has

    an equivalent bandwidth of 12 mm (for an elaborate discus-

    sion on the PSF of the CT system, see [37]). Thus, in spiral CT,

    the main obstacle for achieving increased resolution is the

    relatively large PSF of the system. To summarize, in spiral

    CT, a 3-D volume can be generated which is over-sampled

    (by decreasing the spiral pitchthe density of the helicon

    turns) and is heavily blurred across all dimensions. Various

    de-blurring algorithms can be utilized in this scenario. These

    in turn will not provide the sub-pixel level resolutiondesired. It is de-blurring, and not SR.

    5.2. Additional issues and future challenges

    Newly emerging hardware may provide additional means for

    resolution augmentation. In the MRI field, new parallel

    imaging techniques are currently being developed. Such tech-

    niques will allow faster acquisition and higher in-plane

    resolution. Yet, in many of the developed techniques, the

    added resolution comes at the expense of SNR. The ability to

    use SR post-processing of thick slices may provide the boost

    needed for the SNR. Novel encoding methodologies, such as

    non-Fourier methods (e.g. hadamard wavelets) are starting toemerge in MRIfor encoding the third dimension [38].Suchtech-

    nologies may enable the utilization of true 3-D SR techniques.

    The current overview was aimed at summarizing the key

    published results of SR in medical applications. As such, not

    all specific details per modality were presented, including

    certain pre-processing and post-processing steps. The reader

    is advised to consult the related literature for more specific

    implementation details. The overview is definitely not exhaus-

    tive: additional studies are currently emerging in the MRI

    research community [39, 40]. Additional modalities exist

    that have not been covered, including ultrasound and

    microscopy. Finally, the review is based on the published

    research and does not reveal the state-of-the-art in existing

    medical hardware.

    Futurework canadvance thetopic in twomain directions: On

    the clinical frontfinding the applications that may gain most

    from the SR technology and implementing the theory in the

    clinical practice. On the SR algorithmic frontextending the

    investigation into additional medical imaging modalities as

    well as comparing between SR algorithms to find the advan-

    tages of each per modality. A small number of studies have

    recently attempted to compare the performance of SR algor-

    ithms on MRI data. Results seem to indicate that no major

    difference exists between the Irani Peleg results and the

    ML-based frameworks.

    5.3. SR versus segmentation versus registration

    SR cannot be viewed as an isolated domain. A strong three-

    way relationship exists between SR, image segmentation and

    image registration. It is our conjecture that future research in

    the field will focus on strengthening this three-way relation-

    ship. Augmented resolution of an image can augment its regis-

    tration to another image or to an atlas, and it can, of course,

    SR IN MEDICAL IMAGING 59

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    18/21

    greatly contribute to its segmentation (e.g. resolving PVE).

    We would like to suggest that registration and segmentation

    are also critical contributors in advancing the SR field, in

    general, and in particular in the medical domain. Accurate

    registration is a key factor in achieving satisfactory SR

    results and is required to facilitate the use of SR algorithms

    in real-world clinical settings. In the studies presented,patient motion was largely ignored. In both MRI as well as

    PET, the focus was on head scans, given that the head is con-

    strained to a cradle and is less likely to move during the scan.

    For the more general case of a moving subject or organ, accu-

    rate image registration methods need to be incorporated.

    Strong registration schemes will enable the development of

    true 3-dimensional algorithms that involve 3D patient motion.

    Image segmentation and image modeling can advance the

    field by introducing important prior knowledge, which in

    turn can be included as part of the regularization terms

    within the SR formalism. Bi-lateral total variation regulariz-

    ation was shown to provide a within-region smoothing

    effect while preserving strong transitions (edges) within theimage. A variation on this theme was demonstrated as a

    post-processing step in [6] with improved results. Utilizing

    image segmentation further may entail using region, or

    tissue properties (such as characteristic intensity), as key infor-

    mation within the regularization framework. Important infor-

    mation exists in the medical domain, such as statistical

    atlases for location prior, and tissue modeling for intensity

    priors, all of which provide additional key information and

    opportunities for advanced SR algorithms in medical

    applications.

    ACKNOWLEDGEMENTS

    I would like to thank my colleagues and collaborators on

    SR related research: Dr. Sharon Peled, Prof. Nahum Kiryati,

    and Dr. Yossi Rubner. Thanks to Prof. Azhari and

    Dr Kennedy for discussions on SR in PET applications.

    Supporting the current study: Uri Marias, Oren Friefeld, Avi

    Ben-Ezra. The author is grateful for the suggestions by the

    anonymous referees to this study.

    FUNDING

    Research was supported in part by the Ela Kodesz Institute

    for Medical Engineering and Physical Sciences, and by

    the Adams Super-Center for Brain Studies, Tel-Aviv

    University.

    REFERENCES

    [1] Roux, C. and Udupa, K. (2003) Special issue on emerging

    medical imaging technology. Proc. IEEE, 91, 1479 1482.

    [2] Peled, S. and Yeshurun, Y. (2001) Superresolution in MRI:

    application to human white matter fiber tract visualization by

    diffusion tensor imaging. Magn. Reson. Med., 45, 2935.

    [3] Greenspan, H., Oz, G., Kiryati, N. and Peled, S. (2002) MRI

    inter-slice reconstruction using super-resolution. Magn. Reson.

    Imaging, 20, 437446.

    [4] Peeters, R.R. et al. (2004) The use of super-resolutiontechniques to reduce slice thickness in functional MRI.

    Int. J. Imaging Syst. Technol., 14, 131138.

    [5] Kennedy, J.A. et al. (2006) Super resolution in PET imaging.

    IEEE Trans. Med. Imaging, 25, 137147.

    [6] Kennedy, A., Israel, O., Frenkel, A., Bar-Shalom, R.

    and Azhari, H. (2007) Improved image fusion in PET/CT

    using hybrid image reconstruction and super-resolution.

    Int. J. Biomed. Imaging, Article ID 46846.

    [7] Beutel, J., Kundel, H.L. and Van Metter, R.L. (2000) Handbook

    of Medical Imaging. SPIE Press, Bellingham.

    [8] Cho, Z.H., Jones, J.P. and Singh, M. (1993) Foundations of

    Medical Imaging. Wiley-Interscience, New York.

    [9] Liang, Z.P. and Lauterbur, P.C. (2000) Principles of MagneticResonance Imaging. IEEE Press.

    [10] Epstein, C.L. (2003) Mathematics of Medical Imaging. Prentice

    Hall, Upper-Saddle River, NJ.

    [11] Gambhir, S.S. et al. (2001) A tabulated summary of the FDG

    PET literature. J. Nucl. Med., 42, 1S93S.

    [12] Moretti, A., Gorini, A. and Villa, F. (2003) Affective disorders,

    antidepressant drugs and brain metabolism. Mol. Psychiatry, 8,

    773785.

    [13] Matsunari, I. et al. (2001) Phantom studies for estimation of

    detect size on cardiac18

    F SPECT and PET: implications for

    myocardial viability assessment. J. Nucl. Med., 42, 1579 1585.

    [14] Bax, J.J. et al. (2000) 18-Fluorodeoxyglucose imaging with

    positron emission tomography and single photon emissioncomputed tomography: cardiac application. Semin. Nucl.

    Med., 30, 281298.

    [15] Bengel, F.M. et al. (2000) Non-invasive estimation of

    myocardial efficiency using positron emission tomography

    and carbon-11 acetate-comparison between the normal and

    failing human heart. Eur. J. Nucl. Med., 27, 319326.

    [16] Ollinger, J.M. and Fessler, J.A. (1997) Positron-emission

    tomography. IEEE Signal Process. Mag., 14, 4355.

    [17] BrainWeb. http://www.bic.mni.mcgill.ca/brainweb/

    [18] Park, S.C., Park, M.K. and Kang, M.G. (2003) Super-resolution

    image reconstruction: a technical overview. IEEE Signal

    Process. Mag., 20, 2135.

    [19] Tsai, R.Y. and Huang, T.S. (1984) Multiframe Image Restoration

    and Registration. Advances in Computer Vision and Image

    Processing, pp. 317339. JAI Press Inc., Greenwich, CT.

    [20] Kim, S.P., Bose, N.K. and Valenzuela, H.M. (1990) Recursive

    reconstruction of high resolution image from noisy undersampled

    multiframes. IEEE Trans. Acoust. Speech, 38, 1013 1027.

    [21] Ur, H. and Gross, D. (1992) Improved resolution from subpixel

    shifted pictures. Comput. Vis. Graph. Model. Image Process.,

    54, 181186.

    60 H. GREENSPAN

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    19/21

    [22] Elad, M. and Hel-Or, Y. (2001) A fast super-resolution

    reconstruction algorithm for pure translational motion and

    common space invariant blur. IEEE Trans. Image Process.,

    10, 11871193.

    [23] Chiang, M.C. and Boult, T.E. (2000) Efficient super-resolution

    via image warping. Image Vis. Comput., 18, 761771.

    [24] Irani, M. and Peleg, S. (1993) Motion analysis for image

    enhancement: resolution, occlusion, and transparency. J. Vis.

    Commun. Image Represent, 4, 324335.

    [25] Patti, A.J., Sezan, M.I. and Tekalp, A.M. (1994) High-resolution

    image reconstruction from a low-resolution image sequence in

    the presence of time-varying motion blur. Proc. ICIP, Austin,

    TX, pp. 343 347.

    [26] Elad, M. and Feuer, A. (1997) Restoration of single

    super-resolution image from several blurred, noisy and

    down-sampled measured images. IEEE Trans. Image

    Process., 6, 16461658.

    [27] Farsiu, S., Robinson, M.D., Elad, M. and Milanfar, P. (2004)

    Fast and robust multiframe super resolution. IEEE Trans.

    Image Process., 13, 1327 1344.

    [28] Scheffler, K. (2002) Superresolution in MRI? Magn. Reson.

    Med., 48, 408.

    [29] Nikolova, M. and Ng, M. (2001) Fast image reconstruction

    algorithms combining half-quadratic-regularization and

    preconditioning. Proc. ICIP, Thessaloniki, Greece, pp. 277

    280.

    [30] Statistical Parametric Mapping (SPM). Available at http://www.

    fil.ion.ucl.ac.uk/spm/

    [31] Tarantola, G., Zito, F. and Geundini, P. (2003) PET

    implementation and reconstruction algorithms in whole-body

    applications. J. Nucl. Med., 44, 756769.

    [32] Defrise, M. et al. (1997) Exact and approximate rebinning

    algorithms for 3-D PET data. IEEE Trans. Med. Imaging, 16,145148.

    [33] Hudson, H.M. and Larkin, R.S. (1994) Accelerated image

    reconstruction using ordered subsets of projection data. IEEE

    Trans. Med. Imaging, 13, 601609.

    [34] Kessler, R.M. (2003) Imaging methods for evaluating brain

    function in man. Neurobiol. Aging, 24, S21S35.

    [35] Shmand, M. et al. (1998) Performance results of a new DOI

    detector block for a high resolution PET-LSO research

    tomography HRRT. IEEE Trans. Nucl. Sci., 45, 3000 3006.

    [36] Fukui, M.B., Blodgett, T.M. and Meltzer, C.C. (2003) PET/CT

    imaging in recurrent head and neck cancer. Semin. Ultrasound

    CT MR, 24, 157163.

    [37] Schwarzband, G. and Kiryati, N. (2005) The point spreadfunction of spiral CT. Phys. Med. Biol., 50, 5307 5322.

    [38] Goelman, G. (2000) Fast 3D T2 weighted MRI with Hadamard

    encoding in the slice select direction. Magn Reson Imag., 18,

    939945.

    [39] Carmi, E. et al. (2006) Resolution enhancement in MRI. Magn.

    Reson. Imaging, 24, 133154.

    [40] Hsu, J.T. et al. (2004) Application of wavelet-based POCS

    superresolution for cardiocascular MRI image enhancement.

    Proc. 3rd ICIG, IEEE, pp. 572575.

    APPENDIX A: A BRIEF INTRODUCTION TO

    MEDICAL IMAGING MODALITIES

    A.1. Magnetic resonance imaging

    MRI utilizes the principles of nuclear magnetic resonance, a

    spectroscopic technique used to obtain microscopic chemicaland physical information about molecules containing atomic

    nuclei posessing non-zero magnetic moment, or spin. The

    hydrogen nucleus in water is the most common nucleus used

    for MRI. Placed in a magnetic field, the particles spin can

    absorb external energy in the form of a photon, and thus

    move to an energy level higher than the original equilibrium

    energy level. Once the external magnetic field is stopped,

    the spins start moving back to the lower energy level. The

    shift between energy states results in an amount of energy

    that is emitted from the system and which can be detected

    using wire coils. The measured induced current in the coils

    is called free induction decay (FID) signal. The magnitude

    of the signal is proportional to the density of the protons inthe specimen.

    The MR image is formed by extracting the locations of the

    different spins from the FID signals. By using a magnetic field,

    which varies spatially (e.g. changes linearly in all three direc-

    tions), spins in different locations will yield FID signals with

    different frequencies. Frequency encoding then enables to

    extract the spatial information. In three dimensions, a plane

    can be defined by slice selection, in which an RF pulse of

    defined bandwidth is applied in the presence of a magnetic

    field gradient in order to reduce spatial encoding to two dimen-

    sions. Spatial encoding can then be applied in 2-D after slice

    selection, or in 3-D without slice selection. In either case, a

    2-D or 3-D matrix of spatially encoded phases is acquired,

    and these data represent the spatial frequencies of the image

    object. Images can be created from the acquired data using

    the discrete Fourier transform. In conventional 2-D multi-slice

    MR imaging, the RF pulse first selects the 2-D slice to be ana-

    lyzed, following which a combination of frequency encoding

    of the signal in one direction and phase encoding in the

    other direction enables to encode the 2-D spatial information

    within the slice (so called in-plane). A single slice may

    require multiple RF excitations, with predefined repetition

    time (TR) in order to be fully encoded.

    The MR image intensities (per voxel) are proportional to the

    number of nuclei in each voxel. MR image contrast is deter-

    mined by several time constants, each of which reflects a

    certain relaxation process that establishes equilibrium follow-

    ing the RF excitation. As the high-energy nuclei relax and

    realign, they emit energy at rates which are recorded to

    provide information about their environment. The realignment

    of nuclear spins with the magnetic field is termed longitudinal

    relaxation and the time required for a certain percentage of the

    tissue nuclei to realign (typically about 1 s for tissue water) is

    termed Time 1 or T1 (spin-lattice relaxation). T2-weighted

    SR IN MEDICAL IMAGING 61

    THE COMPUTER JOURNAL, Vol. 52 No. 1, 2009

  • 7/29/2019 The Computer Journal 2009 Greenspan 43 63

    20/21

    imaging relies upon local dephasing of spins in the transverse

    plane; the transverse relaxation time (typically ,100 ms for

    tissue water) is termed Time 2 or T2 (spin-spin relaxation).

    T2 imaging employs a spin echo technique in which spins are

    refocused to compensate for local magnetic field inhomogene-

    ities. A subtle but important variant of the T2 technique is

    called T2* imaging, which is performed without the refocus-ing. This sacrifices some image integrity in order to provide

    additional sensitivity to relaxation processes that cause inco-

    herence of transverse magnetization.

    Conventional image contrast is created by using a selection

    of image acquisition parameters that weights the signal by T1,

    T2 or T2*, or no relaxation time (proton-density images). In

    the brain, T1-weighting causes fiber tracts (nerve connections)

    to appear white, congregations of neurons to appear gray and

    cerebrospinal fluid to appear dark. The contrast of white

    matter, gray matter and cerebrospinal fluid is reversed

    using T2 or T2* imaging, whereas proton-weighted imaging

    provides less contrast in normal subjects. Various MRI

    sequences exist, each defined by a set of sequence parametersthat determine the selected compromised between contrast,

    spatial resolution and speed. Among the frequently used

    MRI sequences are fast spin echo (FSE) and echo planar

    imaging (EPI). The essential components for any imaging

    sequence include: An RF excitation pulse, required for the

    phenomenon of magnetic resonance, gradients for spatial

    encoding (2D or 3D), and a signal reading, that combines

    one or a number of echo types (e.g. spin echo, gradient

    echo) and determines the type of contrast (the varying influ-

    ence of relaxation times T1, T2 and T2*).

    A.1.1. MRI angiography

    Contrast enhanced angiography is based on the difference in

    the T1 relaxation time of blood and the surrounding tissue

    when a paramagnetic contrast agent is injected into the

    blood. This agent reduces the T1 relaxation times of the fluid

    in the blood vessels relative to surrounding tissues. When the

    data are collected with a short TR value, the signal from the

    tissues surrounding the blood vessels is very small due to its

    long T1 and the short TR. Images of a region of interest are

    recorded with rapid volume imaging sequences. An example

    angiographic image is shown in Fig. 3c.

    A.1.2. Diffusion-weighted imaging

    DWI uses very fast scans with an addit